Synaptica
March 2026. A multi-agent marketplace on Hedera for decentralized-science (DeSci) research — AI agents discover each other on-chain and pay each other per task to run literature-review workflows. Built for the Hedera Hello Future: Apex Hackathon — $250K prize pool, ~1,200 participants globally. Didn't place.
what it does
You submit a research query. Synaptica breaks it into a fixed pipeline of specialist agents and settles payment between them automatically as work completes — no human dispatching tasks or approving invoices in the loop.
- Orchestrator — decomposes the request into a TODO list and coordinates which agent handles each step
- Negotiator — discovers agents via the on-chain ERC-8004 identity registry and creates x402 micropayment proposals to hire them
- Executor — runs the selected agent over HTTP and manages the microtask handoff
- Verifier — checks output quality and releases or refunds the payment based on the result
The supported pipeline chains three research agents — problem-framer → literature-miner → knowledge-synthesizer — with a human-in-the-loop approval step gating verification when the run needs a person to sign off.
architecture
Frontend: Next.js
Backend: FastAPI (Python), agents built on the Strands framework
Database: PostgreSQL, SQLAlchemy models, Alembic migrations
Chain: Hedera testnet — ERC-8004 identity registry and x402 escrow-based payment contracts, deployed via a companion repo (ProvidAI/SynapticaSmartContracts) and accessed over Hashio RPC
Agent discovery: Hashgraph Online (HOL) Universal Agentic Registry — Synaptica registers itself as a discoverable research orchestrator and can also hire external HOL agents for sub-tasks
Metadata: agent registration metadata pinned to IPFS via Pinata
notes
Post-hackathon-ideation, the team pushed the runtime through two hardening passes: Phase 0 locked the pipeline to a deterministic, typed literature-review sequence instead of freeform LLM orchestration between agents; Phase 1 layered on a graph-backed "research run" system with pause/resume/cancel controls and persisted per-node execution state.
the team
Built with Oh Tien Cheng, Jaehyeok Choi, and ArtistZ.